Time-Series Growth Prediction Model Based on U-Net and Machine Learning in Arabidopsis

Area of Science:

• Agricultural Science
• Plant Biology
• Computational Biology

Background:

• High-throughput phenotyping platforms (HTPP) generate extensive plant data, but its use in yield prediction is limited.
• Challenges include a lack of quality image analysis, associated yield data, and effective time-series analysis methods.

Purpose of the Study:

• To develop a robust method for crop yield prediction using HTPP data.
• To overcome limitations in image analysis and time-series modeling for phenotyping data.
• To establish associations between early plant development and final yield.

Main Methods:

• Utilized deep learning (DL) networks (U-Net with SE-ResXt101 encoder) for feature extraction from plant images.
• Employed machine learning (ML) algorithms (XGBoost) to identify critical time intervals for prediction.
• Analyzed time-series image data from *Arabidopsis* over 23 days post-sowing.

Main Results:

• Accurate prediction of late pre-flowering yield (23 DAS) using early-stage data (17-21 DAS) was achieved (P < 0.01).
• Projected Area (PA) was successfully estimated into Fresh Weight (FW) with a correlation coefficient of 0.85.
• Demonstrated the feasibility of predicting FW across different developmental stages using time-series analysis.

Conclusions:

• This study pioneers the use of time-series HTPP data for predicting crop yield and biomass.
• The findings offer valuable insights for understanding leafy plant yield and vertical farming biomass.
• Highlights the potential for reducing data complexity and expanding time-series analysis in HTPPs.